Bio-Inspired Quadruped Spiderbot

2025 | Links: Final Report (PDF) | Isaac Lab Training | ROS2 Workspace

A low-cost 12-DOF quadruped robot with forward kinematics gait control, integrated SLAM, and YOLOv8 object detection. Parallel simulation studies exploring CPG-RL approaches using Van der Pol oscillators.

A ** 12-DOF quadruped spiderbot** with two parallel development tracks: a deployed teleoperated system using forward kinematics with predefined gait patterns, integrated with LiDAR-based SLAM and YOLOv8 object detection; and simulation studies exploring Central Pattern Generator-Reinforcement Learning (CPG-RL) approaches using Van der Pol oscillators for bio-inspired locomotion control.

Stack: ROS2 Humble · Arduino Mega · Raspberry Pi 4 · YDLiDAR X2 · MPU9250 IMU · DS3225 Servos · Isaac Lab · RSL-RL (PPO) · Van der Pol CPG · SLAM Toolbox · YOLOv8.

What’s in this project?

Deployed System

12-DOF quadruped platform with custom 3D-printed mechanical design
Forward kinematics-based gait control with predefined crawling patterns
6-DOF motion commands: forward, backward, lateral, and yaw movements
LiDAR-based SLAM for real-time mapping and localization
YOLOv8 object detection for environment awareness
Teleoperated control over ROS2 network

Simulation Studies

CPG-RL architecture learning to modulate Van der Pol oscillator parameters
Comparison studies: Pure RL, Hopf CPG, and Van der Pol CPG approaches
4096 parallel environments in NVIDIA Isaac Lab
Multi-task training with domain randomization
Sim-to-real transfer challenges documented

Gallery

Hardware architecture: Raspberry Pi 4, Arduino Mega, PCA9685 drivers, and sensor integration.

System Design

Mechanical Architecture

The quadruped spiderbot features a symmetric configuration with four legs, each consisting of three joints:

Coxa (hip rotation): Controls leg orientation
Femur (thigh elevation): Controls leg lift
Tibia (knee flexion): Controls foot placement

Total: 12 degrees of freedom

Design Challenges & Solutions

1. Structural Design:

Initial designs cracked under component weight at joint locations
Solution: Increased infill percentage and switched to gyroid infill pattern for improved strength-to-weight ratio
Complete robot resize required to accommodate all components

2. Actuation System:

DS3225 digital servos (25 kg·cm torque @ 6V) for each joint
Two PCA9685 16-channel PWM drivers via I2C bus
Challenges encountered:
- PWM calibration critical for accurate angle mapping
- Excessive current draw during static poses
- Motor disengagement under excessive load
- Careful load distribution required for stability

3. Power System:

18V 10Ah lithium-ion drill battery (stripped down)
Power distribution via three buck converters:
- Two 20A buck converters (5V): Power six servos each
- One 5A buck converter (5V): Powers Raspberry Pi 4

CAD Assembly Video

CAD assembly process showing the 3D-printed components and servo integration.

Electronics and Computing Architecture

Hierarchical Control System:

1. Low-Level Control (Arduino Mega 2560):

Real-time servo control and IMU data acquisition
Gait engine with predefined joint angle sequences
PWM signal generation through PCA9685 drivers
MPU9250 IMU data collection and streaming
Serial command interface for motion commands

2. High-Level Control (Raspberry Pi 4):

Ubuntu Server 22.04 headless with ROS2 Humble
YDLiDAR X2 data acquisition
USB camera feed processing
ROS2 node communication with remote workstation
Serial communication with Arduino for gait commands
ROS_DOMAIN_ID=7 for network communication

Deployed System: Forward Kinematics Gait Control

Kinematics Approach

Due to the complexity of inverse kinematics for a 12-DOF system with joint offsets, the deployed system uses forward kinematics only. Each leg is modeled as a 3-DOF kinematic chain, and foot positions are computed from joint angles to verify reachable workspace and joint limits.

Gait Implementation

Gaits are implemented as predefined sequences of joint angles stored in lookup tables:

Crawling gait: 75% duty factor with predefined swing/stance phases
Motion commands: Forward, backward, left, right, yaw-left, yaw-right
Smooth transitions: Timed interpolation between gait sequences

SLAM Integration

ROS2 Nodes:

ydlidar_ros2_driver: Publishes /scan topic
slam_toolbox: Performs online SLAM, publishing /map
robot_state_publisher: Broadcasts TF transforms

Object Detection

YOLOv8 model processes the USB camera feed for real-time object detection, enabling environment awareness during teleoperation.

Hardware Demo: Walking on Ground

Spiderbot executing forward kinematics-based crawling gait with 75% duty factor.

Simulation Studies: CPG-RL Approaches

In parallel with the deployed system, we conducted simulation experiments exploring learning-based locomotion control in NVIDIA Isaac Lab with 4096 parallel environments.

Pure Reinforcement Learning (Initial Attempt)

First approach: Direct joint control with PPO

Action space: 12D continuous (direct joint positions)
Observations: IMU data (no joint encoders)

Problems encountered:

Unnatural joint angles beyond comfortable ranges
Jerky, discontinuous motion
“Cheating” behaviors (dragging body, hopping on two legs)
Complete failure during hardware deployment attempts

This motivated exploration of structured approaches.

CPG-Based Approaches

Hopf Oscillators

Initial CPG implementation used Hopf oscillators, but the purely sinusoidal waveforms produced overly smooth, unnatural gaits without distinct swing/stance phases.

Van der Pol Oscillators

Following prior work, we implemented Van der Pol oscillators: \[\ddot{x} - \mu(1 - (x/A)^2)\dot{x} + \omega^2 x = 0\]

where:

$\mu = 2.5$ controls nonlinearity
$A$ is amplitude
$\omega = 2\pi f$ is angular frequency

Advantages:

Non-sinusoidal waveforms with distinct swing/stance phases
Adjustable duty cycle via $\mu$ parameter
Sharp phase transitions

Hybrid CPG-RL Architecture

The final simulation architecture has the RL policy learn CPG parameters rather than direct joint commands:

Action Space (17D):

Frequency (1D): 0.3–2.5 Hz (shared)
Amplitudes (12D): 0.0–0.7 rad (per joint)
Phase offsets (4D): ±1.0 rad (per leg)

Observation Space:

Actor (52D): Commands, history, IMU data, CPG phases, previous actions
Critic (76D): Actor observations + privileged joint state (training only)

This asymmetric actor-critic design allows the actor to be deployed without joint encoders while the critic uses privileged information during training.

Training Results

Using RSL-RL (PPO implementation), training over 4096 parallel environments produced:

Smooth forward, lateral, and yaw motions in simulation
Multi-task training (random commands) outperformed curriculum learning
Domain randomization applied: mass ±30%, friction ±25%, actuator parameters ±25%

Isaac Sim Locomotion Videos

Forward Motion:

Lateral & Yaw Motion:

Yaw Rotation (Slow):

Sim-to-Real Challenges

Initial deployment attempts failed due to:

Challenge	Impact
Heavy mass	3.2 kg approaches servo torque limits
Servo dynamics	Compliance, backlash, and latency not modeled
Friction mismatch	Simulation $\mu=1.0$ vs. real $\mu\approx0.6$
CPG frequency	Required reduction from 2.0 Hz to 1.3 Hz
IMU noise	MPU9250 drift and noise not matched in simulation
Open-loop control	No joint encoders to measure actual positions

Sim-to-real transfer of the CPG-RL policy remains future work.

Results

Deployed System Performance

The forward kinematics teleoperated system successfully demonstrates:

✅ Stable crawling gait execution
✅ 6-DOF motion commands (forward, backward, lateral, yaw)
✅ Real-time SLAM map generation
✅ Object detection via YOLOv8
✅ Reliable remote operation over ROS2

Simulation Results Summary

Approach	Motion Quality	Stability	Sim-to-Real
Pure RL	Unnatural	Poor	Failed
Hopf CPG	Smooth	Moderate	Not tested
VDP CPG-RL	Natural	Good	In progress

Challenges and Lessons Learned

Mechanical Design

Component weight must be considered from initial design
Gyroid infill patterns significantly improve joint strength
Iterative prototyping essential for proper component housing

Servo Motor Challenges

PWM-to-angle calibration critical for multi-DOF coordination
Static current draw can exceed driver ratings
Load distribution prevents motor disengagement

Control System Insights

Forward kinematics sufficient when IK complexity exceeds implementation resources
Predefined gaits provide reliable baseline for teleoperation
Open-loop control fundamentally limits precision without encoders

Sim-to-Real Gap

Hobby servos behave significantly differently from simulated actuators
Domain randomization alone insufficient for low-cost hardware
Sensor-minimal designs require careful IMU noise modeling

Future Work

Complete sim-to-real transfer of CPG-RL policy
Integration of joint encoders for closed-loop control
Extension to rough terrain locomotion
Autonomous navigation combining SLAM with learned locomotion
Improved servo dynamics modeling in simulation
Enhanced domain randomization for low-cost hardware

Code Availability

The source code for this project is publicly available:

Isaac Lab Training: spiderbot - Contains simulation environment, CPG-RL training configurations, and trained policies.
ROS2 Workspace: spiderbot.ws - Contains ROS2 packages for robot control, SLAM integration, teleoperation, and sensor processing.

Acknowledgments

This work was conducted at the School for Engineering of Matter, Transport and Energy at Arizona State University under the supervision of Dr. Hamidreza Marvi.

← Back to Projects